One AI improvement in your business. Every month.
No six-figure statement of work. No four-month discovery phase. A flat monthly fee, one working improvement shipped into your operations every month, and a public log of everything I've shipped so you can see the pace before you pay for it.
Everything I've shipped, in order, including what didn't work
Updated the first working day of each month. Client names anonymised, results not.
Three of fourteen were killed. That ratio is the point of working monthly — a bad idea costs one month and gets switched off, instead of being defended for a year because it was in the contract.
The project model is why most AI work disappoints
Not because the engineering is hard. Because a four-month project has to guess what's valuable on day one, and then defend that guess to the end.
- Scope agreed before anyone knows what works
- Four months before anything reaches a real user
- Change requests when reality disagrees with the spec
- A bad idea gets built anyway, because it's in the contract
- Six figures committed on a guess
- Handover at the end, if there's budget left
- We pick the next thing together, monthly, from evidence
- Something reaches a real user inside four weeks
- Scope changes are just next month's decision
- A bad idea costs one month and gets switched off
- One month's fee at risk, ever
- Documented as it ships, because you might cancel
The cancel-anytime term isn't generosity — it's the mechanism. If I have to re-earn the retainer every month, I can't afford to spend one on something that doesn't visibly work. That's a much better guarantee than anything I could write into a contract.
What a month actually looks like
Same shape every time, so you always know where things stand without asking.
Pick & measure
Thirty minutes to choose this month's target from the backlog, then I establish the baseline — how long the process takes now, how often it's wrong now. Without that, "improvement" is an opinion.
Build
Heads down. You get a Friday note each week — what's working, what isn't, and whether I still think this is worth finishing. Sometimes the answer is no, and that's a good month too.
Ship & measure again
It goes live to real users, with the before-and-after numbers next to it. Not a demo — a thing your team uses on Monday.
Keep or kill
We look at the numbers together and decide: keep it, keep improving it, or switch it off. Killing something is a normal outcome, not a failure — and it goes in the public log either way.
What I need from you: about two hours a month, mostly from one person who knows the process. Thirty minutes to pick the target, and the rest scattered across the month answering "is this right?" — which nobody else can answer for me.
One number, on the website, no call required
Most people should start at Standard. You can move up or down at any month boundary.
One improvement, one month, no commitment to anything after it.
$4,500One month · credited if you continue
- One shipped improvement
- Baseline and after measurements
- Runs in your infrastructure
- Yours to keep either way
One improvement a month, ongoing, with everything already shipped kept running.
$6,500/moCancel any month · 30 days notice
- One shipped improvement per month
- Maintenance of everything already shipped
- Weekly Friday note, monthly numbers review
- Shared Slack channel, replies same day
- Documentation written as we go
More capacity, and your engineers learning to do this without me.
$12,000/moCancel any month · 30 days notice
- Two to three improvements a month
- In your standups and your tracker
- Paired work with your engineers
- Architecture review on your own AI work
- Explicit goal: you stop needing me
The one promise that matters
If a month ends without something shipped and measured, that month is free. Not discounted, not credited against next month — free, invoiced at zero.
I can offer that because the unit of work is deliberately one month. Nothing I take on is bigger than four weeks, so "we ran out of time" isn't a thing that happens — if a target turns out to be too large, we split it or swap it in week one, before it becomes a problem.
Model and hosting costs are billed by your provider directly, typically $50–$300 a month depending on volume. I don't mark them up.
What people ask before signing up
What counts as "one improvement"?
Something a real person uses that measurably changes how a process runs. Examples from the log above: suggested replies inside a helpdesk, an invoice extraction pipeline, a triage classifier, a search layer over internal documents.
It's deliberately not a fixed unit — some months the improvement is a big new capability, some months it's making last month's thing 15% more accurate, which is often worth more. We agree what it is in week one, so nobody is surprised at the end.
What if we don't know what to improve?
That's normal and it's the first month's job. I sit with the people doing the work, watch how the time actually goes, and come back with three candidates ranked by value and difficulty. You pick.
Most teams are wrong about where their time goes, incidentally — the thing everyone complains about is rarely the thing that costs the most hours.
Isn't a retainer just an excuse to bill forever?
It would be, if you couldn't leave. You can, any month, with thirty days' notice. And the public log means the pace is visible — if the entries slow down or get thin, that's the same as a bad review.
The Embedded tier is explicitly designed to end: the goal there is your engineers doing this without me. I'd rather be a great six-month engagement you tell people about than a two-year one you resent.
Why is everything in our infrastructure?
Because you can cancel any month, and it would be dishonest to offer that while holding your system hostage. Everything runs in your cloud accounts, in your repositories, using your model API keys. If you leave, nothing switches off and nothing needs migrating.
It also means you can have your own engineers read every line, which is the only real answer to "how do we know it's any good?"
Our data can't leave our network. Is that a problem?
No. Options run from enterprise API tiers with contractual no-training terms and a region you specify, through self-hosted embeddings and vector storage inside your own VPC, to fully local models where nothing crosses your perimeter.
Local models are meaningfully less capable than frontier ones, so if that's your constraint we'll pick targets that suit them. That's a week-one conversation, not a surprise in month three.
What happens if you get sick, or take a holiday?
Honestly: the month slips, and per the guarantee, you don't pay for it. That's the real cost of hiring one person instead of an agency, and I'd rather state it than bury it.
What I do control: two clients at a time, so there's slack in the schedule, and everything documented as it ships rather than at the end — so a pause is a pause, not a crisis.
Can we just hire someone instead?
If you have enough of this work to fill a role, you should — a full-time engineer is better value than any retainer, and I'll say so on the call.
This shape makes sense when there's real work but not a full role's worth, when you want to find out whether there's a role's worth before committing to a salary, or when you need someone senior enough to know what not to build and can't justify that seniority full-time.
One engineer, eight years of things that had to keep running
I'm Bahman Shadmehr. Before this I spent eight years on production backends where quiet failure was expensive: automated trading systems placing orders on the Texas power market, a national payment gateway I helped break out of its monolith, and the Kubernetes clusters and data pipelines underneath both.
That's most of why this offer is shaped around measurement. Shipping something is easy; knowing whether it actually helped, and noticing when it stops helping, is the part that takes experience. Every entry in the log has a number next to it because I don't trust the ones that don't.
I take two clients at a time. When I'm full I say so and give you a date rather than starting badly.
Python, Go, Kubernetes, AWS, GCP, Postgres, Redis, Prometheus. Remote, UTC+3 — European hours and US mornings. 76 posts on DEV · LinkedIn
Tell me what your team did last week that a machine should have done
One paragraph is enough. I'll tell you within a day whether it's a good first month — and if it isn't, what would be.
info@bshadmehr.meUseful in a first message
- What the process is, and who does it.
- Roughly how many hours a week it takes.
- Where the information it needs lives today.
- Whether your data can leave your network.
- Whether you have engineers who'd want to be involved.